Computer Science

Colloquium Details

Scheduling Network Resources in Grid Computing Environments

Abstract

Computational Grids are an exciting new paradigm in distributed computing. Due to the dynamic nature of Grid environments, careful scheduling of programs must be done to insure acceptable performance. This scheduling often uses forecasts of resource performance to adapt runtime scheduling. This talk considers considers mechanisms by which network and service performance can be improved in the context of Grid applications. The first approach to improving Grid performance focuses on end-to-end network throughput. By introducing schedulable "depots" in the network, we are able to increase bandwidth available to hosts communicating over using high-speed, wide-area networks. We describe this technique and present empirical results from our implementation. The next approach deals with statistical prediction techniques for network performance. Performance forecasting for networks has a unique set of constraints, and we introduce and evaluate a mechanism for addressing these. Finally, we address the information system responsible for delivering performance forecasts to clients making runtime scheduling decisions. This portion of the system must scale to large numbers of resources without performance degradation. We introduce a caching proxy capable of client-oriented data aggregation and evaluate the performance improvements.

Biography

Martin Swany is a Ph.D. candidate at the University of California, Santa Barbara. He received a B.A. and M.S. from the University of Tennessee in 1992 and 1998, respectively. His research interests include Grid and distributed computing with an emphasis on enabling network technologies and services.